path: root/compiler/luci-interpreter/src/kernels/Conv2D.cpp

/*
 * Copyright (c) 2020 Samsung Electronics Co., Ltd. All Rights Reserved
 * Copyright 2019 The TensorFlow Authors. All Rights Reserved.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *    http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */

#include "kernels/Conv2D.h"

#include "kernels/Utils.h"

#include <tensorflow/lite/kernels/internal/optimized/legacy_optimized_ops.h>

#include <stdexcept>
#include <thread>

namespace luci_interpreter
{
namespace kernels
{

Conv2D::Conv2D(const Tensor *input, const Tensor *filter, const Tensor *bias, Tensor *output,
               const Conv2DParams &params)
  : KernelWithParams<Conv2DParams>({input, filter, bias}, {output}, params)
{
}

void Conv2D::configure()
{
  // TensorFlow Lite (as of v2.2.0) supports the following combinations of types:
  //     | input filter bias  output |
  // ----+---------------------------+
  // (1) | float float  float float  |
  // (2) | float int8   float float  | hybrid
  // (3) | uint8 uint8  int32 uint8  | quantized
  // (4) | int8  int8   int32 int8   | quantized per channel
  //
  // We only support (1) and (3) for now, and additionally the following:
  //     | input filter bias  output |
  // ----+---------------------------+
  // (5) | int16 int16  int64 int16  |
  //
  if (input()->element_type() == DataType::FLOAT32 && filter()->element_type() == DataType::FLOAT32)
  {
    LUCI_INTERPRETER_CHECK(bias() == nullptr || bias()->element_type() == DataType::FLOAT32);
  }
  else if (input()->element_type() == DataType::U8 && filter()->element_type() == DataType::U8)
  {
    LUCI_INTERPRETER_CHECK(bias() == nullptr || bias()->element_type() == DataType::S32);
  }
  else if (input()->element_type() == DataType::S16 && filter()->element_type() == DataType::S16)
  {
    LUCI_INTERPRETER_CHECK(bias() == nullptr || bias()->element_type() == DataType::S64);
  }
  else
  {
    throw std::runtime_error("Unsupported type.");
  }
  LUCI_INTERPRETER_CHECK(output()->element_type() == input()->element_type());

  const Shape &input_shape = input()->shape();
  const Shape &filter_shape = filter()->shape();
  LUCI_INTERPRETER_CHECK(input_shape.num_dims() == 4 && filter_shape.num_dims() == 4);

  const int32_t batches = input_shape.dim(0);
  const int32_t input_height = input_shape.dim(1);
  const int32_t input_width = input_shape.dim(2);
  const int32_t output_depth = filter_shape.dim(0);
  const int32_t filter_height = filter_shape.dim(1);
  const int32_t filter_width = filter_shape.dim(2);
  LUCI_INTERPRETER_CHECK(filter_shape.dim(3) == input_shape.dim(3));

  LUCI_INTERPRETER_CHECK(bias() == nullptr || (bias()->shape().num_dims() == 1 &&
                                               bias()->shape().dim(0) == output_depth));

  const int32_t output_height =
    computeOutputSize(_params.padding, input_height, filter_height, _params.stride_height,
                      _params.dilation_height_factor);
  const int32_t output_width =
    computeOutputSize(_params.padding, input_width, filter_width, _params.stride_width,
                      _params.dilation_width_factor);

  _padding_height = computePadding(_params.stride_height, _params.dilation_height_factor,
                                   input_height, filter_height, output_height);
  _padding_width = computePadding(_params.stride_width, _params.dilation_width_factor, input_width,
                                  filter_width, output_width);

  output()->resize({batches, output_height, output_width, output_depth});

  // Allocate tensor for Im2Col, if needed.
  // The checks here should be aligned with the actual implementation.
  const bool need_dilated_im2col =
    _params.dilation_height_factor != 1 || _params.dilation_width_factor != 1;
  const bool need_non_dilated_im2col = _params.stride_height != 1 || _params.stride_width != 1 ||
                                       filter_height != 1 || filter_width != 1;
  const bool need_im2col =
    input()->element_type() != DataType::S16 && (need_dilated_im2col || need_non_dilated_im2col);
  if (need_im2col)
  {
    const int input_depth = input_shape.dim(3);
    Shape im2col_shape{batches, output_height, output_width,
                       input_depth * filter_height * filter_width};
    try
    {
      _im2col =
        std::make_unique<Tensor>(input()->element_type(), im2col_shape, AffineQuantization{}, "");
    }
    catch (std::bad_alloc &ba)
    {
      // Failed memory allocation
      _im2col = nullptr;
    }
  }
}

void Conv2D::execute() const
{
  switch (input()->element_type())
  {
    case DataType::FLOAT32:
      if (filter()->element_type() == DataType::FLOAT32)
      {
        evalFloat();
        break;
      }
      throw std::runtime_error("Unsupported type.");
    case DataType::U8:
      if (filter()->scales().size() == 1)
      {
        evalQuantized();
      }
      else if (filter()->scales().size() > 1)
      {
        LUCI_INTERPRETER_CHECK(filter()->shape().num_dims() == 4);
        LUCI_INTERPRETER_CHECK(filter()->scales().size() ==
                               static_cast<size_t>(filter()->shape().dim(0)));
        evalQuantizedPerChannel();
      }
      break;
    case DataType::S16:
      evalQuantizedS16();
      break;
    default:
      throw std::runtime_error("Unsupported type.");
  }
  if (!!_im2col)
    _im2col->deallocate();
}

void Conv2D::evalFloat() const
{
  float activation_min{};
  float activation_max{};
  calculateActivationRange(_params.activation, &activation_min, &activation_max);

  tflite::ConvParams params{};
  params.padding_values.height = _padding_height;
  params.padding_values.width = _padding_width;
  params.stride_height = _params.stride_height;
  params.stride_width = _params.stride_width;
  params.dilation_height_factor = _params.dilation_height_factor;
  params.dilation_width_factor = _params.dilation_width_factor;
  params.float_activation_min = activation_min;
  params.float_activation_max = activation_max;

  if (_im2col)
  {
    try
    {
      tflite::optimized_ops::Conv(
        params, getTensorShape(input()), getTensorData<float>(input()), getTensorShape(filter()),
        getTensorData<float>(filter()), getTensorShape(bias()), getTensorData<float>(bias()),
        getTensorShape(output()), getTensorData<float>(output()), getTensorShape(_im2col.get()),
        getTensorData<float>(_im2col.get()));
    }
    catch (std::bad_alloc &ba)
    {
      // Failed memory allocation
      _im2col->deallocate();

      tflite::reference_ops::Conv(
        params, getTensorShape(input()), getTensorData<float>(input()), getTensorShape(filter()),
        getTensorData<float>(filter()), getTensorShape(bias()), getTensorData<float>(bias()),
        getTensorShape(output()), getTensorData<float>(output()), tflite::RuntimeShape(), nullptr);
    }
  }
  else
    tflite::reference_ops::Conv(
      params, getTensorShape(input()), getTensorData<float>(input()), getTensorShape(filter()),
      getTensorData<float>(filter()), getTensorShape(bias()), getTensorData<float>(bias()),
      getTensorShape(output()), getTensorData<float>(output()), tflite::RuntimeShape(), nullptr);
}

void Conv2D::evalQuantized() const
{
  const auto input_scale = static_cast<double>(input()->scale());
  const auto filter_scale = static_cast<double>(filter()->scale());
  const auto output_scale = static_cast<double>(output()->scale());

  const double real_multiplier = input_scale * filter_scale / output_scale;
  int32_t output_multiplier{};
  int output_shift{};
  quantizeMultiplier(real_multiplier, &output_multiplier, &output_shift);

  int32_t activation_min{};
  int32_t activation_max{};
  calculateActivationRangeQuantized(_params.activation, output(), &activation_min, &activation_max);

  tflite::ConvParams params{};
  params.padding_values.height = _padding_height;
  params.padding_values.width = _padding_width;
  params.stride_height = _params.stride_height;
  params.stride_width = _params.stride_width;
  params.dilation_height_factor = _params.dilation_height_factor;
  params.dilation_width_factor = _params.dilation_width_factor;
  // The kernel expects input and filter zero points to be negated.
  params.input_offset = -input()->zero_point();    // Note the '-'.
  params.weights_offset = -filter()->zero_point(); // Note the '-'.
  params.output_offset = output()->zero_point();
  params.output_multiplier = output_multiplier;
  params.output_shift = output_shift;
  params.quantized_activation_min = activation_min;
  params.quantized_activation_max = activation_max;

  // TODO This should only be done once (although it takes only a few microseconds).
  //  Also, the user should be able to adjust the number of threads.
  auto gemmlowp_context = std::make_unique<gemmlowp::GemmContext>();
  gemmlowp_context->set_max_num_threads(static_cast<int>(std::thread::hardware_concurrency()));

  tflite::optimized_ops::Conv(
    params, getTensorShape(input()), getTensorData<uint8_t>(input()), getTensorShape(filter()),
    getTensorData<uint8_t>(filter()), getTensorShape(bias()), getTensorData<int32_t>(bias()),
    getTensorShape(output()), getTensorData<uint8_t>(output()), getTensorShape(_im2col.get()),
    getTensorData<uint8_t>(_im2col.get()), gemmlowp_context.get());
}

void Conv2D::evalQuantizedPerChannel() const
{
  const auto *input_data = getTensorData<uint8_t>(input());
  const auto *filter_data = getTensorData<uint8_t>(filter());
  const auto *bias_data = getTensorData<int32_t>(bias());
  auto *output_data = getTensorData<uint8_t>(output());

  const Shape &input_shape = input()->shape();
  const Shape &filter_shape = filter()->shape();
  const Shape &output_shape = output()->shape();

  const int32_t batches = input_shape.dim(0);
  const int32_t input_height = input_shape.dim(1);
  const int32_t input_width = input_shape.dim(2);
  const int32_t input_depth = input_shape.dim(3);
  const int32_t output_depth = filter_shape.dim(0);
  const int32_t filter_height = filter_shape.dim(1);
  const int32_t filter_width = filter_shape.dim(2);
  const int32_t output_height = output_shape.dim(1);
  const int32_t output_width = output_shape.dim(2);

  const int32_t stride_height = _params.stride_height;
  const int32_t stride_width = _params.stride_width;
  const int32_t dilation_height_factor = _params.dilation_height_factor;
  const int32_t dilation_width_factor = _params.dilation_width_factor;

  int32_t activation_min{};
  int32_t activation_max{};
  calculateActivationRangeQuantized(_params.activation, output(), &activation_min, &activation_max);

  const std::vector<double> effective_output_scale =
    getQuantizedConvolutionMultiplers(input()->scale(), filter()->scales(), output()->scale());

  const std::vector<ChannelQuantMultipliers> multipliers_raw =
    quantizeMultipliers(effective_output_scale);
  BroadcastableWrapper<ChannelQuantMultipliers> quant_multipliers(multipliers_raw);

  for (int32_t batch = 0; batch < batches; ++batch)
  {
    for (int32_t out_y = 0; out_y < output_height; ++out_y)
    {
      for (int32_t out_x = 0; out_x < output_width; ++out_x)
      {
        for (int32_t out_c = 0; out_c < output_depth; ++out_c)
        {
          const int32_t in_y_origin = out_y * stride_height - _padding_height;
          const int32_t in_x_origin = out_x * stride_width - _padding_width;
          int32_t acc = 0;
          for (int32_t filter_y = 0; filter_y < filter_height; ++filter_y)
          {
            for (int32_t filter_x = 0; filter_x < filter_width; ++filter_x)
            {
              const int32_t in_y = in_y_origin + dilation_height_factor * filter_y;
              const int32_t in_x = in_x_origin + dilation_width_factor * filter_x;
              if ((in_y >= 0 && in_y < input_height) && (in_x >= 0 && in_x < input_width))
              {
                for (int32_t in_c = 0; in_c < input_depth; ++in_c)
                {
                  const uint8_t input_val =
                    input_data[calcOffset(input_shape, batch, in_y, in_x, in_c)];
                  const uint8_t filter_val =
                    filter_data[calcOffset(filter_shape, out_c, filter_y, filter_x, in_c)];
                  acc += static_cast<int32_t>(input_val - input()->zero_point()) *
                         static_cast<int32_t>(filter_val - filter()->zero_points()[out_c]);
                }
              }
            }
          }
          if (bias_data)
          {
            acc += bias_data[out_c];
          }

          int32_t scaled_acc = tflite::MultiplyByQuantizedMultiplier(
            acc, quant_multipliers[out_c].multiplier, quant_multipliers[out_c].shift);

          scaled_acc += output()->zero_point();
          scaled_acc = std::max(scaled_acc, activation_min);
          scaled_acc = std::min(scaled_acc, activation_max);
          output_data[calcOffset(output_shape, batch, out_y, out_x, out_c)] = scaled_acc;
        }
      }
    }
  }
}

void Conv2D::evalQuantizedS16() const
{
  const auto *input_data = getTensorData<int16_t>(input());
  const auto *filter_data = getTensorData<int16_t>(filter());
  const auto *bias_data = getTensorData<int64_t>(bias());
  auto *output_data = getTensorData<int16_t>(output());

  const Shape &input_shape = input()->shape();
  const Shape &filter_shape = filter()->shape();
  const Shape &output_shape = output()->shape();

  const int32_t batches = input_shape.dim(0);
  const int32_t input_height = input_shape.dim(1);
  const int32_t input_width = input_shape.dim(2);
  const int32_t input_depth = input_shape.dim(3);
  const int32_t output_depth = filter_shape.dim(0);
  const int32_t filter_height = filter_shape.dim(1);
  const int32_t filter_width = filter_shape.dim(2);
  const int32_t output_height = output_shape.dim(1);
  const int32_t output_width = output_shape.dim(2);

  const int32_t stride_height = _params.stride_height;
  const int32_t stride_width = _params.stride_width;
  const int32_t dilation_height_factor = _params.dilation_height_factor;
  const int32_t dilation_width_factor = _params.dilation_width_factor;

  int32_t activation_min{};
  int32_t activation_max{};
  calculateActivationRangeQuantized(_params.activation, output(), &activation_min, &activation_max);

  const std::vector<double> effective_output_scale =
    getQuantizedConvolutionMultiplers(input()->scale(), filter()->scales(), output()->scale());

  const std::vector<ChannelQuantMultipliers> multipliers_raw =
    quantizeMultipliers(effective_output_scale);
  BroadcastableWrapper<ChannelQuantMultipliers> multipliers(multipliers_raw);

  for (int32_t batch = 0; batch < batches; ++batch)
  {
    for (int32_t out_y = 0; out_y < output_height; ++out_y)
    {
      for (int32_t out_x = 0; out_x < output_width; ++out_x)
      {
        for (int32_t out_c = 0; out_c < output_depth; ++out_c)
        {
          const int32_t in_y_origin = out_y * stride_height - _padding_height;
          const int32_t in_x_origin = out_x * stride_width - _padding_width;
          int64_t acc = 0;
          for (int32_t filter_y = 0; filter_y < filter_height; ++filter_y)
          {
            for (int32_t filter_x = 0; filter_x < filter_width; ++filter_x)
            {
              const int32_t in_y = in_y_origin + dilation_height_factor * filter_y;
              const int32_t in_x = in_x_origin + dilation_width_factor * filter_x;
              if ((in_y >= 0 && in_y < input_height) && (in_x >= 0 && in_x < input_width))
              {
                for (int32_t in_c = 0; in_c < input_depth; ++in_c)
                {
                  const int16_t input_val =
                    input_data[calcOffset(input_shape, batch, in_y, in_x, in_c)];
                  const int16_t filter_val =
                    filter_data[calcOffset(filter_shape, out_c, filter_y, filter_x, in_c)];
                  acc += static_cast<int64_t>(input_val) * static_cast<int64_t>(filter_val);
                }
              }
            }
          }
          if (bias_data)
          {
            acc += bias_data[out_c];
          }

          int32_t scaled_acc = tflite::MultiplyByQuantizedMultiplier(
            acc, multipliers[out_c].multiplier, multipliers[out_c].shift);

          scaled_acc = std::max(scaled_acc, activation_min);
          scaled_acc = std::min(scaled_acc, activation_max);

          output_data[calcOffset(output_shape, batch, out_y, out_x, out_c)] = scaled_acc;
        }
      }
    }
  }
}

} // namespace kernels
} // namespace luci_interpreter